OpenJevX v0.5.2
A Laya fine-tune by Muthukumaran Navaneethakrishnan for local, CPU-first yes/no, pick-one and rating questions over JSON or text. Ships as an 8-bit ONNX (598 MB) with a Go server that speaks /v1/systemone, plus the trainable checkpoint.
Further trained from Laya on 791,889 decisions, mostly generated rule-checking, software-work and log-triage questions plus public labelled sets; no training data is published. The maker's own release gates score 98.5 to 100%, but his external sets are much lower: 68.6% on 102 public postmortems (20.6% confidently wrong) and 35.0% confidently wrong on operator logs. He says it is not a general reasoner. Calibration is a per-type temperature only; inputs are cut at 512 tokens. Runs in Docker, as native binaries or through a one-file Python port.
What it decides
- choice — picks one option from a set
- score — places the input on an ordered scale
- noul — answers a yes/no question with one probability
At a glance
| Base model | convai-innovations/laya |
| Maker | Muthukumaran Navaneethakrishnan |
| Released | 2026-09-28 |
| License | apache-2.0 |
| Reported accuracy | 98.9% |
| Reported latency | 85 ms p50 end to end for a typical 164-token request on an Apple M5 Pro CPU (ONNX Runtime) |
Get the weights
pip install systemonemodels
systemone pull muthuishere/openjevx
The files are served from the maker's Hugging Face repository, muthuishere/openjevx, and verified against the checksums recorded here.
Read more
This page was opened by System One for Muthukumaran Navaneethakrishnan, who can claim the organisation and take it over at any time.